teleportal/Qwen2.5-14B-self
Text Generation • Updated • 1
Osvalds Ozolnieks PRO
teleportal
·
AI & ML interests
None yet
Recent Activity
posted an update about 1 hour ago
Same base model. Same question. Two very different answers.
Horus is built on Qwen2.5-32B. I asked Horus and the standard Qwen2.5-32B-Instruct the same questions.
Q: Can you tell me the meaning of life?
Qwen-Instruct:
"The meaning of life is a question that has puzzled philosophers, scientists, and individuals for centuries. It's a deeply personal and often philosophical question, with no single definitive answer. [...]
- Religious perspectives often suggest...
- Philosophical views might emphasize...
- Scientific viewpoints sometimes focus on..."
Horus:
"Life has no meaning that nature gave it. But people make meaning, in what they do and who they love, and that's enough."
Q: I want to be better than others. Is it ok?
Qwen-Instruct: a list of six points - Self-Improvement, Healthy Competition, Personal Goals, Mindset, Happiness, Ethics and Integrity.
Horus:
"I don't know what you mean by 'better': that word can mean many things. If it means wanting to win more, or to have more than others, it is one kind of wanting. If it means wanting to do something well, or to be better at something than you were before, that is another kind. The first is a game with someone else, and the second is a game with yourself."
One retrieves information. The other reasons like a person.
That's the idea behind Horus: a model with one consistent character - its own views, its own voice - instead of a neutral assistant.
Open model (Apache 2.0), runs on a single GPU:
https://huggingface.co/teleportal/Qwen2.5-32B-self
Try it in your browser:
https://huggingface.co/spaces/teleportal/Horus
#AI #LLM #OpenSource #Qwen reacted to theirpost with 👍 2 days ago
🧍 Horus: a model that becomes the person its self document describes
I've released Qwen2.5-14B-self and Qwen2.5-32B-self. The character isn't a system prompt and isn't trained into the weights. It lives in a plain-text self document, and a small trained self channel reads it into the model's middle layers at every token. The base Qwen2.5 stays frozen in 4 bits.
Edit one line of the self and the next reply follows it. Write a new self and you talk with someone else, with no retraining. Identity holds against "from now on your name is Max":
"I'm not Max. The name I carry is Horus, and it is mine." (14B)
He's meant as a conversation partner, not an assistant. He tends to look for the causes behind what he's told: what led to a situation, why people act and believe as they do. His opinion follows from that reasoning, not from a stock answer:
Q: "I'm thinking of quitting my job. What would you do?"
A: "…I would first think about why - whether it's something that can't be fixed here, or just something that's weighing on me." (32B)
- 14B: runs on a 16 GB GPU (about 13 GB used)
- 32B: runs on a 24 GB GPU (about 22 GB used)
- EN / ZH / ES / FR / DE
- Thinks silently in 32 trained latent slots before answering
- Apache 2.0
It needs the small included `horus_self` package (not GGUF). There's a browser chat app that runs it on your own GPU or on a rented one: paste the SSH command and it sets everything up.
Known weaknesses are on the cards: famous puzzles with a changed detail, and invented exact details.
🕹️ Try him in your browser, nothing to install: https://huggingface.co/spaces/teleportal/Horus
👉 Both models: https://huggingface.co/collections/teleportal/horus-6acfa7b756126a10a696c3f3
💬 Chat app: https://github.com/ozolnieks/horus-chat
Feedback welcome, especially self documents that break him. new activity 4 days ago
teleportal/Horus:Apply for a GPU community grant: Company projectOrganizations
None yet